arXiv subjects

Sirisha Rambhatla

Publications and source records attributed to Sirisha Rambhatla.

At least 19 recordsLinked to original sources

arXiv2026

Geometrically Principled Randomized Optimization for Efficient LLM Training

Low-rank gradient optimization for large language models is currently divided into two categories: structured methods that rigorously identify subspaces, and randomized approaches employed primarily for computational efficiency. In this work, we question the intuition behind why random projections are effective. We trace this phenomenon to the geometry of the gradient subspaces, which exhibits subspace optimization landscape has a nearly flat curvature, while a significant portion of gradient information lies outside the core subspace. Leveraging these insights, and drawing on randomized linear algebra, we theoretically establish that random low-rank projections preserve the geometry, and we introduce GrassWalk and GrassJump, algorithms that navigate the Grassmannian manifold via random walks and jumps. By coupling this randomized exploration with subspace-aware optimizer and recovering the lost gradient signals, we achieve state-of-the-art results on LLaMA-1B, LLaMA-7B, and Qwen-1.5B pretraining. Our findings reframe randomization not merely as a computational shortcut, but as a geometrically principled approach to high-dimensional optimizations.

cs.LG↗

arXiv2026

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied datasets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To address this, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks, latent-space representation attacks, and alignment-stage defenses; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. The results provide insights including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that current alignment-stage defenses largely fail to withstand attack sweeps. Code is available at https://github.com/criticalml-uw/TamperBench.

cs.CR↗

arXiv2026

PinPoint: Prompting with Informative Interior Points

Modern referring image segmentation pipelines couple a vision-language model (VLM) for grounding with a promptable segmenter such as the Segment Anything Model (SAM) for mask generation. Prior training-free instances of this recipe consistently trail fine-tuned and reinforcement-learning (RL)-tuned specialists, and it has been unclear whether the gap comes from the VLM's grounding, SAM's capacity, or the prompt. We show that the gap is dominated by prompt ambiguity: a VLM-proposed bounding box (bbox) leaves SAM to guess which pixels inside the bbox belong to the object the expression denotes. Interior points are the natural disambiguator, but where they fall matters; prior work relies on naively sampled points that land on boundaries, distractors, and background clutter, and can even hurt performance compared to the bbox alone. Supervised and RL-tuned methods close this gap by training a VLM to predict better points; we show that this training is unnecessary. At a matched budget of five interior points, replacing naive sampling with stable, informative point selection improves cumulative Intersection-over-Union (cIoU) by 12-18 points across RefCOCO/+/g, with every model fixed. We turn this observation into PinPoint, a deterministic, training-free point selector that fuses four visual cues into a consensus map, selects compact, spatially diverse points away from boundaries, and uses the frozen VLM to label each point. Without any task-specific training, PinPoint matches supervised and RL-tuned specialists on the same stack while issuing only two VLM calls per query.

cs.CV↗

arXiv2026

Zero-Shot Object Re-Identification in Egocentric Kitchen Videos via Multi-Stage SAM3 Feature Fusion

Object re-identification (ReID) in egocentric kitchen videos is challenging due to rapid viewpoint changes, frequent occlusions, cluttered scenes, and large intra-class appearance variations. Objects may leave and re-enter the field of view, and the large diversity of instances with limited annotations makes supervised ReID difficult to scale, motivating zero-shot approaches. We study zero-shot object ReID on the EPIC-Kitchens benchmark, where the goal is to match active food and kitchen-tool instances across frames using only pre-trained visual features. We first evaluate five state-of-the-art feature extractors, including Vision-Language Models (VLMs) - CLIP, DINOv2, DreamSim, I-JEPA, and SAM3 - and show that zero-shot methods fail, with the best baseline achieving only 45.3% mAP. We then propose an Enhanced SAM3 ReID Pipeline, a zero-shot multi-stage method built around SAM3 segmentation as the core component. Stage 1 uses SAM3 to suppress background clutter. Stage 2 fuses embeddings from SAM3, DINOv2, and CLIP into a single L2-normalized descriptor. Stage 3 augments cosine similarity with mask-shape IoU for geometric consistency, and Stage 4 applies k-reciprocal re-ranking. The full pipeline improves performance by 7.5% mAP to 52.8%.

cs.CV↗

arXiv2025

Avatar4D: Synthesizing Domain-Specific 4D Humans for Real-World Pose Estimation

We present Avatar4D, a real-world transferable pipeline for generating customizable synthetic human motion datasets tailored to domain-specific applications. Unlike prior works, which focus on general, everyday motions and offer limited flexibility, our approach provides fine-grained control over body pose, appearance, camera viewpoint, and environmental context, without requiring any manual annotations. To validate the impact of Avatar4D, we focus on sports, where domain-specific human actions and movement patterns pose unique challenges for motion understanding. In this setting, we introduce Syn2Sport, a large-scale synthetic dataset spanning sports, including baseball and ice hockey. Avatar4D features high-fidelity 4D (3D geometry over time) human motion sequences with varying player appearances rendered in diverse environments. We benchmark several state-of-the-art pose estimation models on Syn2Sport and demonstrate their effectiveness for supervised learning, zero-shot transfer to real-world data, and generalization across sports. Furthermore, we evaluate how closely the generated synthetic data aligns with real-world datasets in feature space. Our results highlight the potential of such systems to generate scalable, controllable, and transferable human datasets for diverse domain-specific tasks without relying on domain-specific real data.

cs.CV↗

arXiv2025

SubTrack++ : Gradient Subspace Tracking for Scalable LLM Training

Training large language models (LLMs) is highly resource-intensive due to their massive number of parameters and the overhead of optimizer states. While recent work has aimed to reduce memory consumption, such efforts often entail trade-offs among memory efficiency, training time, and model performance. Yet, true democratization of LLMs requires simultaneous progress across all three dimensions. To this end, we propose SubTrack++ that leverages Grassmannian gradient subspace tracking combined with projection-aware optimizers, enabling Adam's internal statistics to adapt to subspace changes. Additionally, employing recovery scaling, a technique that restores information lost through low-rank projections, further enhances model performance. Our method demonstrates SOTA convergence by exploiting Grassmannian geometry, reducing pre-training wall-time by up to 65% and fine-tuning time by 36% compared to existing SOTA methods, while maintaining the same memory footprint.

cs.LG↗

arXiv2025

Understanding LLM Scientific Reasoning through Promptings and Model's Explanation on the Answers

Large language models (LLMs) have demonstrated remarkable capabilities in natural language understanding, reasoning, and problem-solving across various domains. However, their ability to perform complex, multi-step reasoning task-essential for applications in science, medicine, and law-remains an area of active investigation. This paper examines the reasoning capabilities of contemporary LLMs, analyzing their strengths, limitations, and potential for improvement. The study uses prompt engineering techniques on the Graduate-Level GoogleProof Q&A (GPQA) dataset to assess the scientific reasoning of GPT-4o. Five popular prompt engineering techniques and two tailored promptings were tested: baseline direct answer (zero-shot), chain-of-thought (CoT), zero-shot CoT, self-ask, self-consistency, decomposition, and multipath promptings. Our findings indicate that while LLMs exhibit emergent reasoning abilities, they often rely on pattern recognition rather than true logical inference, leading to inconsistencies in complex problem-solving. The results indicated that self-consistency outperformed the other prompt engineering technique with an accuracy of 52.99%, followed by direct answer (52.23%). Zero-shot CoT (50%) outperformed multipath (48.44%), decomposition (47.77%), self-ask (46.88%), and CoT (43.75%). Self-consistency performed the second worst in explaining the answers. Simple techniques such as direct answer, CoT, and zero-shot CoT have the best scientific reasoning. We propose a research agenda aimed at bridging these gaps by integrating structured reasoning frameworks, hybrid AI approaches, and human-in-the-loop methodologies. By critically evaluating the reasoning mechanisms of LLMs, this paper contributes to the ongoing discourse on the future of artificial general intelligence and the development of more robust, trustworthy AI systems.

cs.AI↗

arXiv2025

Gen4D: Synthesizing Humans and Scenes in the Wild

Lack of input data for in-the-wild activities often results in low performance across various computer vision tasks. This challenge is particularly pronounced in uncommon human-centric domains like sports, where real-world data collection is complex and impractical. While synthetic datasets offer a promising alternative, existing approaches typically suffer from limited diversity in human appearance, motion, and scene composition due to their reliance on rigid asset libraries and hand-crafted rendering pipelines. To address this, we introduce Gen4D, a fully automated pipeline for generating diverse and photorealistic 4D human animations. Gen4D integrates expert-driven motion encoding, prompt-guided avatar generation using diffusion-based Gaussian splatting, and human-aware background synthesis to produce highly varied and lifelike human sequences. Based on Gen4D, we present SportPAL, a large-scale synthetic dataset spanning three sports: baseball, icehockey, and soccer. Together, Gen4D and SportPAL provide a scalable foundation for constructing synthetic datasets tailored to in-the-wild human-centric vision tasks, with no need for manual 3D modeling or scene design.

cs.GR↗

arXiv2025

SafeTuneBed: A Toolkit for Benchmarking LLM Safety Alignment in Fine-Tuning

As large language models (LLMs) become ubiquitous, parameter-efficient fine-tuning methods and safety-first defenses have proliferated rapidly. However, the number of approaches and their recent increase have resulted in diverse evaluations-varied datasets, metrics, and inconsistent threat settings-making it difficult to fairly compare safety, utility, and robustness across methods. To address this, we introduce SafeTuneBed, a benchmark and toolkit unifying fine-tuning and defense evaluation. SafeTuneBed (i) curates a diverse repository of multiple fine-tuning datasets spanning sentiment analysis, question-answering, multi-step reasoning, and open-ended instruction tasks, and allows for the generation of harmful-variant splits; (ii) enables integration of state-of-the-art defenses, including alignment-stage immunization, in-training safeguards, and post-tuning repair; and (iii) provides evaluators for safety (attack success rate, refusal consistency) and utility. Built on Python-first, dataclass-driven configs and plugins, SafeTuneBed requires minimal additional code to specify any fine-tuning regime, defense method, and metric suite, while ensuring end-to-end reproducibility. We showcase its value by benchmarking representative defenses across varied poisoning scenarios and tasks. By standardizing data, code, and metrics, SafeTuneBed is the first focused toolkit of its kind to accelerate rigorous and comparable research in safe LLM fine-tuning. Code is available at: https://github.com/criticalml-uw/SafeTuneBed

cs.LG↗

arXiv2025

Estimating Quality in Therapeutic Conversations: A Multi-Dimensional Natural Language Processing Framework

Engagement between client and therapist is a critical determinant of therapeutic success. We propose a multi-dimensional natural language processing (NLP) framework that objectively classifies engagement quality in counseling sessions based on textual transcripts. Using 253 motivational interviewing transcripts (150 high-quality, 103 low-quality), we extracted 42 features across four domains: conversational dynamics, semantic similarity as topic alignment, sentiment classification, and question detection. Classifiers, including Random Forest (RF), Cat-Boost, and Support Vector Machines (SVM), were hyperparameter tuned and trained using a stratified 5-fold cross-validation and evaluated on a holdout test set. On balanced (non-augmented) data, RF achieved the highest classification accuracy (76.7%), and SVM achieved the highest AUC (85.4%). After SMOTE-Tomek augmentation, performance improved significantly: RF achieved up to 88.9% accuracy, 90.0% F1-score, and 94.6% AUC, while SVM reached 81.1% accuracy, 83.1% F1-score, and 93.6% AUC. The augmented data results reflect the potential of the framework in future larger-scale applications. Feature contribution revealed conversational dynamics and semantic similarity between clients and therapists were among the top contributors, led by words uttered by the client (mean and standard deviation). The framework was robust across the original and augmented datasets and demonstrated consistent improvements in F1 scores and recall. While currently text-based, the framework supports future multimodal extensions (e.g., vocal tone, facial affect) for more holistic assessments. This work introduces a scalable, data-driven method for evaluating engagement quality of the therapy session, offering clinicians real-time feedback to enhance the quality of both virtual and in-person therapeutic interactions.

cs.CL↗

arXiv2025

Human vs. LLM-Based Thematic Analysis for Digital Mental Health Research: Proof-of-Concept Comparative Study

Thematic analysis provides valuable insights into participants' experiences through coding and theme development, but its resource-intensive nature limits its use in large healthcare studies. Large language models (LLMs) can analyze text at scale and identify key content automatically, potentially addressing these challenges. However, their application in mental health interviews needs comparison with traditional human analysis. This study evaluates out-of-the-box and knowledge-base LLM-based thematic analysis against traditional methods using transcripts from a stress-reduction trial with healthcare workers. OpenAI's GPT-4o model was used along with the Role, Instructions, Steps, End-Goal, Narrowing (RISEN) prompt engineering framework and compared to human analysis in Dedoose. Each approach developed codes, noted saturation points, applied codes to excerpts for a subset of participants (n = 20), and synthesized data into themes. Outputs and performance metrics were compared directly. LLMs using the RISEN framework developed deductive parent codes similar to human codes, but humans excelled in inductive child code development and theme synthesis. Knowledge-based LLMs reached coding saturation with fewer transcripts (10-15) than the out-of-the-box model (15-20) and humans (90-99). The out-of-the-box LLM identified a comparable number of excerpts to human researchers, showing strong inter-rater reliability (K = 0.84), though the knowledge-based LLM produced fewer excerpts. Human excerpts were longer and involved multiple codes per excerpt, while LLMs typically applied one code. Overall, LLM-based thematic analysis proved more cost-effective but lacked the depth of human analysis. LLMs can transform qualitative analysis in mental healthcare and clinical research when combined with human oversight to balance participant perspectives and research resources.

cs.HC↗

arXiv2025

LOCATEdit: Graph Laplacian Optimized Cross Attention for Localized Text-Guided Image Editing

Text-guided image editing aims to modify specific regions of an image according to natural language instructions while maintaining the general structure and the background fidelity. Existing methods utilize masks derived from cross-attention maps generated from diffusion models to identify the target regions for modification. However, since cross-attention mechanisms focus on semantic relevance, they struggle to maintain the image integrity. As a result, these methods often lack spatial consistency, leading to editing artifacts and distortions. In this work, we address these limitations and introduce LOCATEdit, which enhances cross-attention maps through a graph-based approach utilizing self-attention-derived patch relationships to maintain smooth, coherent attention across image regions, ensuring that alterations are limited to the designated items while retaining the surrounding structure. LOCATEdit consistently and substantially outperforms existing baselines on PIE-Bench, demonstrating its state-of-the-art performance and effectiveness on various editing tasks. Code can be found on https://github.com/LOCATEdit/LOCATEdit/

cs.CV↗

arXiv2025

Medical Misinformation in AI-Assisted Self-Diagnosis: Development of a Method (EvalPrompt) for Analyzing Large Language Models

Rapid integration of large language models (LLMs) in health care is sparking global discussion about their potential to revolutionize health care quality and accessibility. At a time when improving health care quality and access remains a critical concern for countries worldwide, the ability of these models to pass medical examinations is often cited as a reason to use them for medical training and diagnosis. However, the impact of their inevitable use as a self-diagnostic tool and their role in spreading healthcare misinformation has not been evaluated. This study aims to assess the effectiveness of LLMs, particularly ChatGPT, from the perspective of an individual self-diagnosing to better understand the clarity, correctness, and robustness of the models. We propose the comprehensive testing methodology evaluation of LLM prompts (EvalPrompt). This evaluation methodology uses multiple-choice medical licensing examination questions to evaluate LLM responses. We use open-ended questions to mimic real-world self-diagnosis use cases, and perform sentence dropout to mimic realistic self-diagnosis with missing information. Human evaluators then assess the responses returned by ChatGPT for both experiments for clarity, correctness, and robustness. The results highlight the modest capabilities of LLMs, as their responses are often unclear and inaccurate. As a result, medical advice by LLMs should be cautiously approached. However, evidence suggests that LLMs are steadily improving and could potentially play a role in healthcare systems in the future. To address the issue of medical misinformation, there is a pressing need for the development of a comprehensive self-diagnosis dataset. This dataset could enhance the reliability of LLMs in medical applications by featuring more realistic prompt styles with minimal information across a broader range of medical fields.

cs.CY↗

arXiv2025

LangDA: Building Context-Awareness via Language for Domain Adaptive Semantic Segmentation

Unsupervised domain adaptation for semantic segmentation (DASS) aims to transfer knowledge from a label-rich source domain to a target domain with no labels. Two key approaches in DASS are (1) vision-only approaches using masking or multi-resolution crops, and (2) language-based approaches that use generic class-wise prompts informed by target domain (e.g. "a {snowy} photo of a {class}"). However, the former is susceptible to noisy pseudo-labels that are biased to the source domain. The latter does not fully capture the intricate spatial relationships of objects -- key for dense prediction tasks. To this end, we propose LangDA. LangDA addresses these challenges by, first, learning contextual relationships between objects via VLM-generated scene descriptions (e.g. "a pedestrian is on the sidewalk, and the street is lined with buildings."). Second, LangDA aligns the entire image features with text representation of this context-aware scene caption and learns generalized representations via text. With this, LangDA sets the new state-of-the-art across three DASS benchmarks, outperforming existing methods by 2.6%, 1.4% and 3.9%.

cs.CV↗

arXiv2024

Domain-Guided Masked Autoencoders for Unique Player Identification

Unique player identification is a fundamental module in vision-driven sports analytics. Identifying players from broadcast videos can aid with various downstream tasks such as player assessment, in-game analysis, and broadcast production. However, automatic detection of jersey numbers using deep features is challenging primarily due to: a) motion blur, b) low resolution video feed, and c) occlusions. With their recent success in various vision tasks, masked autoencoders (MAEs) have emerged as a superior alternative to conventional feature extractors. However, most MAEs simply zero-out image patches either randomly or focus on where to mask rather than how to mask. Motivated by human vision, we devise a novel domain-guided masking policy for MAEs termed d-MAE to facilitate robust feature extraction in the presence of motion blur for player identification. We further introduce a new spatio-temporal network leveraging our novel d-MAE for unique player identification. We conduct experiments on three large-scale sports datasets, including a curated baseball dataset, the SoccerNet dataset, and an in-house ice hockey dataset. We preprocess the datasets using an upgraded keyframe identification (KfID) module by focusing on frames containing jersey numbers. Additionally, we propose a keyframe-fusion technique to augment keyframes, preserving spatial and temporal context. Our spatio-temporal network showcases significant improvements, surpassing the current state-of-the-art by 8.58%, 4.29%, and 1.20% in the test set accuracies, respectively. Rigorous ablations highlight the effectiveness of our domain-guided masking approach and the refined KfID module, resulting in performance enhancements of 1.48% and 1.84% respectively, compared to original architectures.

cs.CV↗

arXiv2023

Is Generative Modeling-based Stylization Necessary for Domain Adaptation in Regression Tasks?

Unsupervised domain adaptation (UDA) aims to bridge the gap between source and target domains in the absence of target domain labels using two main techniques: input-level alignment (such as generative modeling and stylization) and feature-level alignment (which matches the distribution of the feature maps, e.g. gradient reversal layers). Motivated from the success of generative modeling for image classification, stylization-based methods were recently proposed for regression tasks, such as pose estimation. However, use of input-level alignment via generative modeling and stylization incur additional overhead and computational complexity which limit their use in real-world DA tasks. To investigate the role of input-level alignment for DA, we ask the following question: Is generative modeling-based stylization necessary for visual domain adaptation in regression? Surprisingly, we find that input-alignment has little effect on regression tasks as compared to classification. Based on these insights, we develop a non-parametric feature-level domain alignment method -- Implicit Stylization (ImSty) -- which results in consistent improvements over SOTA regression task, without the need for computationally intensive stylization and generative modeling. Our work conducts a critical evaluation of the role of generative modeling and stylization, at a time when these are also gaining popularity for domain generalization.

cs.CV↗

arXiv2023

Domain Generalization for Domain-Linked Classes

Domain generalization (DG) focuses on transferring domain-invariant knowledge from multiple source domains (available at train time) to an, a priori, unseen target domain(s). This requires a class to be expressed in multiple domains for the learning algorithm to break the spurious correlations between domain and class. However, in the real-world, classes may often be domain-linked, i.e. expressed only in a specific domain, which leads to extremely poor generalization performance for these classes. In this work, we aim to learn generalizable representations for these domain-linked classes by transferring domain-invariant knowledge from classes expressed in multiple source domains (domain-shared classes). To this end, we introduce this task to the community and propose a Fair and cONtrastive feature-space regularization algorithm for Domain-linked DG, FOND. Rigorous and reproducible experiments with baselines across popular DG tasks demonstrate our method and its variants' ability to accomplish state-of-the-art DG results for domain-linked classes. We also provide practical insights on data conditions that increase domain-linked class generalizability to tackle real-world data scarcity.

cs.LG↗

arXiv2022

Building Spatio-temporal Transformers for Egocentric 3D Pose Estimation

Egocentric 3D human pose estimation (HPE) from images is challenging due to severe self-occlusions and strong distortion introduced by the fish-eye view from the head mounted camera. Although existing works use intermediate heatmap-based representations to counter distortion with some success, addressing self-occlusion remains an open problem. In this work, we leverage information from past frames to guide our self-attention-based 3D HPE estimation procedure -- Ego-STAN. Specifically, we build a spatio-temporal Transformer model that attends to semantically rich convolutional neural network-based feature maps. We also propose feature map tokens: a new set of learnable parameters to attend to these feature maps. Finally, we demonstrate Ego-STAN's superior performance on the xR-EgoPose dataset where it achieves a 30.6% improvement on the overall mean per-joint position error, while leading to a 22% drop in parameters compared to the state-of-the-art.

cs.CV↗